Sakshi Rakesh Kumar

• AI evaluation

  injectbench

  A prompt-injection evaluation harness that reports attack success rate per attack class with Wilson 95% intervals. It uses ablation to measure which guardrail layers actually reduce attack success relative to the latency they add.

  • Five-class attack taxonomy, 400 seeded trials per class (2,000 total)
  • Flags a redundant guardrail layer: 0 of 2000 trials flipped, at an 11 ms cost
  • 44 tests in about 2 s; regression_check usable as a CI gate

  Python · Statistics · JavaScript · Node.js · GitHub Actions/Pages

  SourceLive demo
• Document AI

  fieldextract

  Document field extraction with per-field calibrated confidence. Each field is routed to human review at a cost-optimal threshold, so operations teams know which fields need a human. It runs on a seeded simulated extractor.

  • ECE cut from 0.0290 to 0.0048, which is 6.0x lower, after per-field isotonic calibration
  • Per-field routing costs 1.78x less than one global threshold for the same 4,817 reviews
  • 0 planted PII strings leaked after redaction across 500 docs; 51 tests

  Python · Isotonic calibration · Platt scaling · JavaScript · GitHub Actions/Pages

  SourceLive demo
• Security tooling

  sigmaforge

  A Sigma detection-rule compiler that parses rules into a typed expression tree. It renders the tree as Splunk SPL, Elastic EQL, Microsoft KQL and SQLite SQL. The SQL backend is checked against a reference evaluator with property-based differential tests.

  • Compiles one backend-neutral tree to four query targets
  • 684 tests with mypy --strict and a coverage gate
  • Rules a target cannot express are rejected with a stated reason

  Python 3.12 · SQLite · YAML · Splunk SPL · Elastic EQL · Microsoft KQL · mypy · uv

  Source
• Security testing

  reachable

  A CVE triage tool that checks static call-graph reachability from entry points to the vulnerable function. Each finding is reported as reachable, not-reachable or indeterminate, and the tool never reports a silent 'safe'.

  • Demo app: 2 reachable, 1 indeterminate, 5 not reachable advisories
  • One added getattr() turns all not-reachable verdicts into indeterminate
  • CI gate and SARIF 2.1.0 output; standard library only

  Python · ast · SARIF · GitHub Actions

  Source
• Data quality

  lineagelens

  A column-level SQL lineage and impact analysis tool. It resolves SELECT *, CTEs and UNIONs to show exactly which downstream columns and dashboards break when a source column changes.

  • 10 models touch orders, but only 6 columns and 4 of 5 dashboards read discount_pct
  • Unexpandable stars are surfaced as unresolved, never silently dropped
  • 18 tests plus Node.js tests for the browser demo

  Python · SQL · JavaScript · Node.js · Mermaid · GitHub Pages

  SourceLive demo

More projects analytics, privacy and compliance side projects

• Process analytics

  p2pexceptions

  Procure-to-pay invoice exception analysis that classifies each failed three-way match into one cause from a mutually exclusive, exhaustive taxonomy. The classifier is validated against labelled cases, and causes are ranked in a cost-weighted Pareto.

  • 400,000 synthetic invoice lines, 72,739 failing three-way match (18.2%)
  • 99.6% agreement with a labelled sample of 800; 0.00% unclassified
  • 24 tests covering each precedence rule

  Python · unittest · GitHub Actions

  Source
• Experimentation

  cuped

  A CUPED variance-reduction library for A/B tests that refuses covariates measured after randomisation. Its validation gate tests that bias case explicitly, using a balance check across arms.

  • 38.7% of variance removed; interval width 0.848 to 0.664
  • Gate rejects a post-treatment covariate that returned -6.601 against a true 1.80

  Python · Statistics · unittest · GitHub Actions

  Source
• Compliance automation

  controlmap

  SOC 2 evidence automation that collects exports from GitHub, AWS, identity and HR systems and evaluates YAML controls against them. Controls are mapped to SOC 2 criteria, and the tool builds a hashed evidence package an auditor can verify.

  • 6 built-in collectors and 22 built-in rules
  • Library validated against 61 TSC criteria ids and 93 Annex A controls
  • Evidence package checkable with shasum, with an optional Ed25519 signature

  Python · YAML · SHA-256 · Ed25519 · GitHub Actions

  Source
• Cost modelling

  costtoserve

  An activity-based cost-to-serve model that flags customers whose profitability sign flips depending on which allocation driver is used.

  • Seeded fixture of 200 customers across three segments
  • Seven cost pools, fully reconciled under three driver sets
  • 17 tests; zero leakage under every driver set

  Python

  Source
• Privacy tooling

  privkit

  A privacy toolkit for data pipelines covering PII detection, NIST FF1/FF3-1 format-preserving tokenization, k-anonymity and l-diversity, and differential privacy with a budget accountant.

  • Checksum validators such as Luhn and IBAN mod-97 confirm regex PII candidates
  • Tokens follow NIST FF1 and FF3-1, checked against NIST vectors
  • k-anonymity, l-diversity and t-closeness checks beside a budgeted differential-privacy accountant

  Python

  Source
• Machine learning

  Personality Prediction Project

  A final year project on automated personality classification for e-recruitment. The system analyzes résumés and social media data to predict a candidate's personality, then recommends suitable job profiles.

  • Predicts candidate personality from résumés and social media data
  • Recommends suitable job profiles from the predicted personality
  • Final year project, written in PHP

  PHP · Machine Learning

  Source